Tech leaders to UN: For sake of humanity, please control the AI tech we created
apnews.com
apnews.com
Sam Altman’s remarks at the United Nations Security Council
https://www.youtube.com/watch?v=lHiLFoqNUCg&t=772s (https://openai.com/index/sam-altman-un-security-council-rema...)
Anthropic's Dario Amodei
I think the reasonable thing would have been to let some sort of committee of interested countries to have access to a bunch of relevant. It could still be public and in video form.
I like the call for ensuring that AI isn't concentrated in one company or country though. I would be happy if this were followed, with appropriate EU efforts to make good training accelerators etc. so that we could have our own reasonable 10T+ models.
I've poo-pooed the bio stuff earlier, but I think I agree with the threat to some degree, but I still absolutely hate the idea of central control of models, probably because it's still central control.
No, it is not. Regards, the CIA /s
I guess the parallel to a different industry would be like if some companies called for food safety laws because they think unregulated food is unsafe, and people said “Well those companies who think unregulated food is unsafe can just stop selling food anytime they want, easy peasy”?
In here, I am specifically saying "the specific companies that are producing the unsafe products should stop". That is what people said about food safety too. You are the one poisoning people, you should stop. And crucially, the companies least capable to produce safe food should not be the ones writing the regulations. Their close friends is "independent" organizations should not be the ones writing them either.
Also, effective food safety regulations are about accountability - you poison people, you have to pay. In case of AI, your AI hacks, you are deemed responsible. You pay the fine, you are prosecuted for negligence.
2. Wouldn't that big like only 2 food companies releasing reports when their food makes people sick, then claiming that it's only those 2 companies with bad food, and that all the companies that don't report then their food makes people sick... that those companies don't make people sick? Like if only McDonalds and KFC released reports when their food make people sick, and all the other companies, would you similarly claim that the other companies' food doesn't make people sick?
3.
> In case of AI, your AI hacks, you are deemed responsible. You pay the fine, you are prosecuted
Do you actually believe that is true? I'd say the opposite has been shown to be true - even when it's not a case of "operator gave non-malicious instructions, system acted maliciously" -- i.e. the real-life cases of all of the many government-run and private hacking groups globally for years - do those government-run and private hacking groups actually face consequences for their actions? Or what makes you think things will suddenly change now in terms of consequences?
4. Also, do you think the amount of hacking done with the big frontier models is remotely as significant (either in terms of quantity or how much damage is done) as the amount done with the others?
The initial reporting came from HuggingFace - victim. The labs are trying to shit the blame to "the models" from themselves. We do not have to buy that ridiculous framing.
2.) The situation here is McDonalds and KFC making experiments with putting poop close to food, then reporting the people got sick and then wanting to write rules for other companies.
Nope, first stop putting poop close to food.
3.) Yes, intentional hacking is a crime and yes people went to jail. Yes, CIA and police are known to break the law to get what they want. That does not mean private company should not be responsible for their own law breaking.
4.) What are you even on about here? Are you an LLM that does not understand the context or a troll?
That seems to be misinformation or disinformation - the US labs do also report when others, including nation-state actors, use their models maliciously, agreed?
Or are you saying it's those who aren't for regulation that provide zero transparency around malicious use of their models?
>> do you think the amount of hacking done with the big frontier models is remotely as significant (either in terms of quantity or how much damage is done) as the amount done with the others?
> What are you even on about here?
If you were unable to understand the question, I will attempt to reword it: Do you believe hacking using models from the big US tech labs is anywhere near as prevalent, or causes anywhere near as much damage to its victims, as hacking from other models?
The kind of harms I am talking about are the many trillions of dollars per year in existing cybercrime costs.
When you refer to the OpenAI/Anthropic actions as being like McDonalds/KFC contaminating food with poop, what exactly is the harm that you are referring to? Was it trillions of dollars? Billions? People who committed suicide as a result of it? Or even if you think the harms are trivial or not, what exactly do you think the harms of those hacking incidents were? For reference, the HuggingFace incident involved 17.6K actions on its network - no customer data was leaked, etc.
> The initial reporting came from HuggingFace
HuggingFace knew something had happened but did not have much of an idea beyond that. The only reason we have details about it is exactly because companies like OpenAI provide transparency - if it was instead done by any of the non-regulation groups, we would not have that info, agreed?